Neural speaker diarization is widely used for overlapaware speaker diarization, but it requires large multi-speaker datasets for training. To meet this data requirement, large datasets are often constructed by combining multiple corpora, including those originally designed for multi-speaker automatic speech recognition (ASR). However, ASR datasets often feature loosely defined segment boundaries that do not align with the stricter conventions of diarization benchmarks. In this work, we show that such boundary looseness significantly impacts the diarization error rate, reducing evaluation reliability. We also reveal that models trained on data with varying boundary precision tend to learn dataset-specific looseness, leading to poor generalization across out-of-domain datasets. Training with standardized tight boundaries via forced alignment improves not only diarization performance, especially in streaming scenarios, but also ASR performance when combined with simple post-processing.

This paper proposes a guided speaker embedding extraction system, which extracts speaker embeddings of the target speaker using speech activities of target and interference speakers as clues. Several methods for long-form overlapped multi-speaker audio processing are typically two-staged: i) segment-level processing and ii) inter-segment speaker matching. Speaker embeddings are often used for the latter purpose. Typical speaker embedding extraction approaches only use single-speaker intervals to avoid corrupting the embeddings with speech from interference speakers. However, this often makes speaker embeddings impossible to extract because sufficiently long non-overlapping intervals are not always available. In this paper, we propose using speaker activities as clues to extract the embedding of the speaker-of-interest directly from overlapping speech. Specifically, we concatenate the activity of target and non-target speakers to acoustic features before being fed to the model. We also condition the attention weights used for pooling so that the attention weights of the intervals in which the target speaker is inactive are zero. The effectiveness of the proposed method is demonstrated in speaker verification and speaker diarization.

This paper investigates cascade approaches bridging speech and text foundation models (FMs) for speech translation (ST). We address the limitations of cascade systems which suffer from the propagation of speech recognition errors and the lack of access to acoustic information. We propose a ReShape Attention (RSA) that bridges speech embeddings of Whisper, a speech FM, to LLaMA2, a text FM. Speech and text embeddings have temporal and dimensional gaps, which make merging them challenging. RSA reshapes the speech and text embeddings into a sequence of subvectors sharing the same feature dimension. RSA performs cross-attention in the LLaMA2 layers between these two sequences, which allows combining the two embeddings. The RSA allows text FM to directly access speech FM embeddings and optimize the entire ST system for input speech. RSA improves 8.5% relative BLEU score compared to the baseline ST system, which cascades Whisper and LLaMA2. Moreover, our analyses show that the proposed method could even improve performance with ground-truth transcriptions, which suggests that our bridging approach is not limited to mitigating the effect of recognition errors but can also exploit the benefit of acoustic information.

Extending the RNN Transducer (RNNT) to recognize multi-talker speech is essential for wider automatic speech recognition (ASR) applications. Multi-talker RNNT (MT-RNNT) aims to achieve recognition without relying on costly front-end source separation. MT-RNNT is conventionally implemented using architectures with multiple encoders or decoders, or by serializing all speakers’ transcriptions into a single output stream. The first approach is computationally expensive, particularly due to the need for multiple encoder processing. In contrast, the second approach involves a complex label generation process, requiring accurate timestamps of all words spoken by all speakers in the mixture, obtained from an external ASR system. In this paper, we propose a novel alignment-free training scheme for the MT-RNNT (MT-RNNT-AFT) that adopts the standard RNNT architecture. The target labels are created by appending a prompt token corresponding to each speaker at the beginning of the transcription, reflecting the order of each speaker’s appearance in the mixtures. Thus, MT-RNNT-AFT can be trained without relying on accurate alignments, and it can recognize all speakers’ speech with just one round of encoder processing. Experiments show that MT-RNNT-AFT achieves performance comparable to that of the state-of-the-art alternatives, while greatly simplifying the training process.

Self-supervised learning (SSL) models have significantly advanced speech processing tasks, and several benchmarks have been proposed to validate their effectiveness. However, previous benchmarks have primarily focused on single-speaker scenarios, with less exploration of target-speaker tasks in noisy, multi-talker conditions—a more challenging yet practical case. In this paper, we introduce the Target-Speaker Speech Processing Universal Performance Benchmark (TS-SUPERB), which includes four widely recognized target-speaker processing tasks that require identifying the target speaker and extracting information from the speech mixture. In our benchmark, the speaker embedding extracted from enrollment speech is used as a clue to condition downstream models. The benchmark result reveals the importance of evaluating SSL models in target speaker scenarios, demonstrating that performance cannot be easily inferred from related single-speaker tasks. Moreover, by using a unified SSL-based target speech encoder, consisting of a speaker encoder and an extractor module, we also investigate joint optimization across TS tasks to leverage mutual information and demonstrate its effectiveness.1

Self-supervised learning (SSL) has attracted increased attention for learning meaningful speech representations. Speech SSL models, such as WavLM, employ masked prediction training to encode general-purpose representations. In contrast, speaker SSL models, exemplified by DINO-based models, adopt utterance-level training objectives primarily for speaker representation. Understanding how these models represent information is essential for refining model efficiency and effectiveness. Unlike the various analyses of speech SSL, there has been limited investigation into what information speaker SSL captures and how its representation differs from speech SSL or other fully-supervised speaker models. This paper addresses these fundamental questions. We explore the capacity to capture various speech properties by applying SUPERB evaluation probing tasks to speech and speaker SSL models. We also examine which layers are predominantly utilized for each task to identify differences in how speech is represented. Furthermore, we conduct direct comparisons to measure the similarities between layers within and across models. Our analysis unveils that 1) the capacity to represent content information is somewhat unrelated to enhanced speaker representation, 2) specific layers of speech SSL models would be partly specialized in capturing linguistic information, and 3) speaker SSL models tend to disregard linguistic information but exhibit more sophisticated speaker representation.

The zero-shot text-to-speech (TTS) method, based on speaker embeddings extracted from reference speech using self-supervised learning (SSL) speech representations, can reproduce speaker characteristics very accurately. However, this approach suffers from degradation in speech synthesis quality when the reference speech contains noise. In this paper, we propose a noise-robust zero-shot TTS method. We incorporated adapters into the SSL model, which we fine-tuned with the TTS model using noisy reference speech. In addition, to further improve performance, we adopted a speech enhancement (SE) front-end. With these improvements, our proposed SSL-based zero-shot TTS achieved high-quality speech synthesis with noisy reference speech. Through the objective and subjective evaluations, we confirmed that the proposed method is highly robust to noise in reference speech, and effectively works in combination with SE.

Target-speaker speech processing (TS) tasks, such as target-speaker automatic speech recognition (TS-ASR), target speech extraction (TSE), and personal voice activity detection (p-VAD), are important for extracting information about a desired speaker’s speech even when it is corrupted by interfering speakers. While most studies have focused on training schemes or system architectures for each specific task, the auxiliary network for embedding target-speaker cues has not been investigated comprehensively in a unified crosstask evaluation. Therefore, this paper aims to address a fundamental question: what is the preferred speaker embedding for TS tasks? To this end, for the TS-ASR, TSE, and p-VAD tasks, we compare pre-trained speaker encoders (i.e., self-supervised or speaker recognition models) that compute speaker embeddings from pre-recorded enrollment speech of the target speaker with ideal speaker embeddings derived directly from the target speaker’s identity in the form of a one-hot vector. To further understand the properties of ideal speaker embedding, we optimize it using a gradient-based approach to improve performance on the TS task. Our analysis reveals that speaker verification performance is somewhat unrelated to TS task performances, the one-hot vector outperforms enrollment-based ones, and the optimal embedding depends on the input mixture.

This paper proposes a method for extracting speaker embedding for each speaker from a variable-length recording containing multiple speakers. Speaker embeddings are crucial not only for speaker recognition but also for various multi-speaker speech applications such as speaker diarization and target-speaker speech processing. Despite the challenges of obtaining a single speaker’s speech without pre-registration in multi-speaker scenarios, most studies on speaker embedding extraction focus on extracting embeddings only from single-speaker recordings. Some methods have been proposed for extracting speaker embeddings directly from multi-speaker recordings, but they typically require preparing a model for each possible number of speakers or involve complicated training procedures. The proposed method computes the embeddings of multiple speakers by focusing on different parts of the frame-wise embeddings extracted from the input multi-speaker audio. This is achieved by recursively computing attention weights for pooling the frame-wise embeddings. Additionally, we propose using the calculated attention weights to estimate the number of speakers in the recording, which allows the same model to be applied to various numbers of speakers. Experimental evaluations demonstrate the effectiveness of the proposed method in speaker verification and diarization tasks.

Recently, multi-decoder and universal models have attracted increased interest in speech and language processing as they allow learning common representations across tasks. These models learn a common representation by sharing a part of or all network parameters. Moreover, such a universal model can handle tasks unseen during training (zero-shot tasks). However, these models do not fully exploit inter-dependencies between tasks during decoding since they usually perform decoding for each task independently. In this paper, we propose to address this issue by extending the universal model to perform multi-task parallel decoding with a cross-attention module between decoders to capture task inter-dependencies explicitly. We also introduce a novel multi-stream beam search algorithm to allow such parallel decoding. We test our proposed model on multi-lingual (English and Portuguese) text/speech translation and summarization, confirming its potential, especially in zero-shot tasks.

Self-supervised learning (SSL) has been dramatically successful not only in monolingual but also in cross-lingual settings. However, since the two settings have been studied individually in general, there has been little research focusing on how effective a cross-lingual model is in comparison with a monolingual model. In this paper, we investigate this fundamental question empirically with Japanese automatic speech recognition (ASR) tasks. First, we begin by comparing the ASR performance of cross-lingual and monolingual models for two different language tasks while keeping the acoustic domain as identical as possible. Then, we examine how much unlabeled data collected in Japanese is needed to achieve performance comparable to a cross-lingual model pre-trained with tens of thousands of hours of English and/or multilingual data. Finally, we extensively investigate the effectiveness of SSL in Japanese and demonstrate state-of-the-art performance on multiple ASR tasks. Since there is no comprehensive SSL study for Japanese, we hope this study will guide Japanese SSL research.

End-to-end speech summarization (E2E SSum) is a technique to directly generate summary sentences from speech. Compared with the cascade approach, which combines automatic speech recognition (ASR) and text summarization models, the E2E approach is more promising because it mitigates ASR errors, incorporates nonverbal information, and simplifies the overall system. However, since collecting a large amount of paired data (i.e., speech and summary) is difficult, the training data is usually insufficient to train a robust E2E SSum system. In this paper, we present two novel methods that leverage a large amount of external text summarization data for E2E SSum training. The first technique is to utilize a text-to-speech (TTS) system to generate synthesized speech, which is used for E2E SSum training with the text summary. The second is a TTS-free method that directly inputs phoneme sequence instead of synthesized speech to the E2E SSum model. Experiments show that our proposed TTS- and phoneme-based methods improve several metrics on the How2 dataset. In particular, our best system outperforms a previous state-of-the-art one by a large margin (i.e., METEOR score improvements of more than 6 points). To the best of our knowledge, this is the first work to use external language resources for E2E SSum. Moreover, we report a detailed analysis of the How2 dataset to confirm the validity of our proposed E2E SSum system.

The recurrent neural network-transducer (RNNT) is a promising approach for automatic speech recognition (ASR) with the introduction of a prediction network that autoregressively considers linguistic aspects. To train the autoregressive part, the ground-truth tokens are used as substitutions for the previous output token, which leads to insufficient robustness to incorrect past tokens; a recognition error in the decoding leads to further errors. Scheduled sampling (SS) is a technique to train autoregressive model robustly to past errors by randomly replacing some ground-truth tokens with actual outputs generated from a model. SS mitigates the gaps between training and decoding steps, known as exposure bias, and it is often used for attentional encoder-decoder training. However SS has not been fully examined for RNNT because of the difficulty in applying SS to RNNT due to the complicated RNNT output form. In this paper we propose SS approaches suited for RNNT. Our SS approaches sample the tokens generated from the distiribution of RNNT itself, i.e. internal language model or RNNT outputs. Experiments in three datasets confirm that RNNT trained with our SS approach achieves the best ASR performance. In particular, on a Japanese ASR task, our best system outperforms the previous state-of-the-art alternative.

In this paper, we propose novel cross-lingual self-supervised speech representation learning methods that explicitly consider language information. Cross-lingual self-supervised speech representation learning has been studied to make effective use of diverse data in various languages. Previous methods train models from multilingual datasets without taking language into account. However, it is difficult to train speech representations from multilingual datasets in the same space without language specification since there are clear differences in the acoustic context between languages. To solve this problem, we propose leveraging language IDs to build self-supervised speech representation learning models that explicitly consider language information. Our proposed models utilize fixed-dimensional language embeddings converted from language IDs for the model learning the relationship between related speech representations in different languages. We investigate two strategies to introduce language embeddings into the models: adding the embeddings to all of the inputs and concatenating to the inputs of the Transformer. We experimentally investigated how the difference between the two strategies affects the downstream tasks. Experimental results on the English and Japanese datasets show that the proposed methods improve the accuracies of downstream automatic speech recognition tasks.

We propose a new Customer Satisfaction Estimation (CSE) method that utilizes unsupervised representation learning. Though conventional methods have improved both the heuristic features and the estimation models, their performance is still insufficient as only small amounts of labeled training data can be expected. To mitigate this problem, the proposed method leverages a large amount of unlabeled data by unsupervised representation learning based on self-training. The key advance of the proposed method is to introduce a Multi-Format Prediction (MFP) loss to improve the performance of self-training for the inputs that contain both continuous and biased discrete features such as the number of occurrences of a particular word. MFP loss uses two loss functions based on regression and weighted binary classification to reconstruct both types of features with high accuracy. Experiments on real English contact center calls reveal the improved CSE performance attained by the proposed method.

In this paper we propose improvements to our recently proposed hybrid RNN-T/Attention architecture that includes a shared encoder followed by recurrent neural network-transducer (RNN-T) and triggered attention-based decoders (TAD). The use of triggered attention enables the attention-based decoder (AD) to operate in a streaming manner. When a trigger point is detected by RNN-T, TAD uses the context from the start-of-speech up to that trigger point to compute the attention weights. Consequently, the computation costs and the memory consumptions are quadratically increased with the duration of the utterances because all input features must be stored and used to re-compute the attention weights. In this paper, we use a short context from a few frames prior to each trigger point for attention weight computation resulting in reduced computation and memory costs. We call the proposed framework triggered chunkwise AD (TCAD). We also investigate the effectiveness of internal language model (ILM) estimation approach using both ILMs of RNN-T and TCAD heads for improving RNN-T performance. We confirm in experiments with public and private datasets covering various scenarios that TCAD achieves superior recognition performance while reducing computation costs compared to TAD.

This paper investigates the effectiveness and implementation of modality-specific large-scale pre-trained encoders for multimodal sentiment analysis (MSA). Although the effectiveness of pre-trained encoders in various fields has been reported, conventional MSA methods employ them for only linguistic modality, and their application has not been investigated. This paper compares the features yielded by large-scale pre-trained encoders with conventional heuristic features. One each of the largest pre-trained encoders publicly available for each modality are used; CLIP-ViT, WavLM, and BERT for visual, acoustic, and linguistic modalities, respectively. Experiments on two datasets reveal that methods with domain-specific pre-trained encoders attain better performance than those with conventional features in both unimodal and multimodal scenarios. We also find it better to use the outputs of the intermediate layers of the encoders than those of the output layer. The codes are available at https://github.com/ando-hub/MSA_Pretrain.

This paper presents a novel speech emotion recognition scheme that can deal with the individuality of emotion perception. Most conventional methods directly model the majority decision of multiple listener’s perceived emotions. However, emotion perception varies with the listener, which means the conventional methods can mismatch the recognition results to human perception. In order to mitigate this problem, we propose a Listener Adaptive (LA) model that reflects emotion recognition criteria of each listener. One-hot listener codes with several adaptation layers are employed in the LA model. The LA model yields the posterior probabilities of the listener-specific perceived emotions. Majority-voted emotion can be also estimated by averaging, in the LA model, the posterior probabilities for all listeners. Experiments on two emotional speech datasets demonstrate that the proposed approach offers improved listener-wise perceived emotion recognition performance in natural speech.

Recurrent neural network-transducer (RNN-T) is promising for building time-synchronous end-to-end automatic speech recognition (ASR) systems, in part because it does not need frame-wise alignment between input features and target labels in the training step. Although training without alignment is beneficial, it makes it difficult to discern the relation between input features and output token sequences. This, in effect, degrades RNN-T performance. Our solution is SimpleFlat (SF), a novel and simple whole-network pretraining approach for RNN-T. SF extracts frame-wise alignments on-the-fly from the training dataset, and does not require any external resources. We distribute equal numbers of target tokens to each frame following RNN-T encoder output lengths by repeating each token. The frame-wise tokens so created are shifted, and also used as the prediction network inputs. Therefore, SF can be implemented by cross entropy loss computation as in autoregressive model training. Experiments on Japanese and English ASR tasks demonstrate that SF can effectively improve various RNN-T architectures.

We propose a cross-modal transformer-based neural correction models that refines the output of an automatic speech recognition (ASR) system so as to exclude ASR errors.Generally, neural correction models are composed of encoder-decoder networks, which can directly model sequence-to-sequence mapping problems.The most successful method is to use both input speech and its ASR output text as the input contexts for the encoder-decoder networks.However, the conventional method cannot take into account the relationships between these two different modal inputs because the input contexts are separately encoded for each modal.To effectively leverage the correlated information between the two different modal inputs, our proposed models encode two different contexts jointly on the basis of cross-modal self-attention using a transformer.We expect that cross-modal self-attention can effectively capture the relationships between two different modals for refining ASR hypotheses.We also introduce a shallow fusion technique to efficiently integrate the first-pass ASR model and our proposed neural correction model.Experiments on Japanese natural language ASR tasks demonstrated that our proposed models achieve better ASR performance than conventional neural correction models.

We present a novel training approach for connectionist temporal classification (CTC) -based automatic speech recognition (ASR) systems. CTC models are promising for building both a conventional acoustic model and an end-to-end (E2E) ASR model. However, CTC models make it difficult to capture the correct timing of each output label because timing is not given explicitly in the training data. In this paper, we propose a new auxiliary task with frame-wise targets for CTC model enhancement. We utilize attention weights generated by an attention-based encoder-decoder model (S2S) for making the targets, called the attention matrix. The attention matrix is the sum of the products of the attention weights (spike timing information) and the corresponding target vectors (probability information), and used for S2S-to-CTC knowledge distillation loss computation. Therefore, the attention matrix makes the CTC models jointly train-able as regards spike timings and their posteriors. Experiments on Japanese ASR tasks demonstrate that our proposal is effective for CTC model training; it achieves a 10.2% (E2E) / 9.4% (acoustic model) relative reduction in the character/kana-syllable error rates compared to models trained using only CTC loss.